def build_model(opt, start_from):
if opt.att_model == 'topdown':
model = AttModel.TopDownModel(opt)
else:
model = AttModel.Att2in2Model(opt)
if start_from is not None:
if opt.load_best_score == 1:
model_path = os.path.join(start_from, 'model-best.pth')
info_path = os.path.join(start_from,
'infos_' + opt.id + '-best.pkl')
else:
model_path = os.path.join(start_from, 'model.pth')
info_path = os.path.join(start_from, 'infos_' + opt.id + '.pkl')
# opt.learning_rate = saved_model_opt.learning_rate
print('Loading the model weights, path is %s...' % (model_path))
model.load_state_dict(torch.load(model_path))
if opt.mGPUs:
model = nn.DataParallel(model)
if opt.cuda:
model.cuda()
return model
opt.fg_mask = torch.from_numpy(dataset.fg_mask).byte()
opt.glove_fg = torch.from_numpy(dataset.glove_fg).float()
opt.glove_clss = torch.from_numpy(dataset.glove_clss).float()
opt.glove_w = torch.from_numpy(dataset.glove_w).float()
opt.st2towidx = torch.from_numpy(dataset.st2towidx).long()
opt.itow = dataset.itow
opt.itod = dataset.itod
opt.ltow = dataset.ltow
opt.itoc = dataset.itoc
if not opt.finetune_cnn:
opt.fixed_block = 4 # if not finetune, fix all cnn block
if opt.att_model == 'topdown':
model = AttModel.TopDownModel(opt)
elif opt.att_model == 'att2in2':
model = AttModel.Att2in2Model(opt)
tf_summary_writer = tf and tf.summary.FileWriter(opt.checkpoint_path)
infos = {}
histories = {}
if opt.start_from is not None:
if opt.load_best_score == 1:
model_path = os.path.join(opt.start_from, 'model-best.pth')
info_path = os.path.join(opt.start_from,
'infos_' + opt.id + '-best.pkl')
else:
model_path = os.path.join(opt.start_from, 'model.pth')
info_path = os.path.join(opt.start_from,
'infos_' + opt.id + '.pkl')
def init_model(opt):
cudnn.benchmark = True
####################################################################################
# Build the Model
####################################################################################
from misc.dataloader_coco import DataLoader
dataset = DataLoader(opt, split='train')
opt.vocab_size = dataset.vocab_size
opt.detect_size = dataset.detect_size
opt.seq_length = opt.seq_length
opt.fg_size = dataset.fg_size
opt.fg_mask = torch.from_numpy(dataset.fg_mask).byte()
opt.glove_fg = torch.from_numpy(dataset.glove_fg).float()
opt.glove_clss = torch.from_numpy(dataset.glove_clss).float()
opt.glove_w = torch.from_numpy(dataset.glove_w).float()
opt.st2towidx = torch.from_numpy(dataset.st2towidx).long()
opt.itow = dataset.itow
opt.itod = dataset.itod
opt.ltow = dataset.ltow
opt.itoc = dataset.itoc
if not opt.finetune_cnn: opt.fixed_block = 4 # if not finetune, fix all cnn block
if opt.att_model == 'topdown':
model = AttModel.TopDownModel(opt)
elif opt.att_model == 'att2in2':
model = AttModel.Att2in2Model(opt)
infos = {}
histories = {}
if opt.start_from is not None:
if opt.load_best_score == 1:
model_path = os.path.join(opt.start_from, 'model-best.pth')
info_path = os.path.join(opt.start_from, 'infos_'+opt.id+'-best.pkl')
else:
model_path = os.path.join(opt.start_from, 'model.pth')
info_path = os.path.join(opt.start_from, 'infos_'+opt.id+'.pkl')
# open old infos and check if models are compatible
with open(info_path, 'rb') as f:
infos = pickle.load(f, encoding='latin1')
saved_model_opt = infos['opt']
# opt.learning_rate = saved_model_opt.learning_rate
print('Loading the model %s...' %(model_path))
model.load_state_dict(torch.load(model_path))
if os.path.isfile(os.path.join(opt.start_from, 'histories_'+opt.id+'.pkl')):
with open(os.path.join(opt.start_from, 'histories_'+opt.id+'.pkl'), 'rb') as f:
histories = pickle.load(f, encoding='latin1')
if opt.decode_noc:
model._reinit_word_weight(opt, dataset.ctoi, dataset.wtoi)
best_val_score = infos.get('best_val_score', None)
iteration = infos.get('iter', 0)
start_epoch = infos.get('epoch', 0)
val_result_history = histories.get('val_result_history', {})
loss_history = histories.get('loss_history', {})
lr_history = histories.get('lr_history', {})
ss_prob_history = histories.get('ss_prob_history', {})
if opt.cuda:
model.cuda()
params = []
# cnn_params = []
for key, value in dict(model.named_parameters()).items():
if value.requires_grad:
if 'cnn' in key:
params += [{'params':[value], 'lr':opt.cnn_learning_rate,
'weight_decay':opt.cnn_weight_decay, 'betas':(opt.cnn_optim_alpha, opt.cnn_optim_beta)}]
else:
params += [{'params':[value], 'lr':opt.learning_rate,
'weight_decay':opt.weight_decay, 'betas':(opt.optim_alpha, opt.optim_beta)}]
print("Use %s as optmization method" %(opt.optim))
if opt.optim == 'sgd':
optimizer = optim.SGD(params, momentum=0.9)
elif opt.optim == 'adam':
optimizer = optim.Adam(params)
elif opt.optim == 'adamax':
optimizer = optim.Adamax(params)
obj_det_model = prep_model('../faster-rcnn.pytorch/data_fr/pretrained_model/faster_rcnn_coco.pth')
return obj_det_model, model
def eval_nbt(opt):
cudnn.benchmark = True
if opt.dataset == 'flickr30k':
from misc.dataloader_flickr30k import DataLoader
else:
from misc.dataloader_coco import DataLoader
####################################################################################
# Data Loader
####################################################################################
dataset_val = DataLoader(opt, split=opt.val_split)
dataloader_val = torch.utils.data.DataLoader(dataset_val,
batch_size=1,
shuffle=False,
num_workers=1)
input_imgs = torch.FloatTensor(1)
input_seqs = torch.LongTensor(1)
input_ppls = torch.FloatTensor(1)
gt_bboxs = torch.FloatTensor(1)
mask_bboxs = torch.ByteTensor(1)
gt_seqs = torch.LongTensor(1)
input_num = torch.LongTensor(1)
if opt.cuda:
input_imgs = input_imgs.cuda()
input_seqs = input_seqs.cuda()
gt_seqs = gt_seqs.cuda()
input_num = input_num.cuda()
input_ppls = input_ppls.cuda()
gt_bboxs = gt_bboxs.cuda()
mask_bboxs = mask_bboxs.cuda()
input_imgs = Variable(input_imgs)
input_seqs = Variable(input_seqs)
gt_seqs = Variable(gt_seqs)
input_num = Variable(input_num)
input_ppls = Variable(input_ppls)
gt_bboxs = Variable(gt_bboxs)
mask_bboxs = Variable(mask_bboxs)
####################################################################################
# Build the Model
####################################################################################
opt.vocab_size = dataset_val.vocab_size
opt.detect_size = dataset_val.detect_size
opt.seq_length = opt.seq_length
opt.fg_size = dataset_val.fg_size
opt.fg_mask = torch.from_numpy(dataset_val.fg_mask).byte()
opt.glove_fg = torch.from_numpy(dataset_val.glove_fg).float()
opt.glove_clss = torch.from_numpy(dataset_val.glove_clss).float()
opt.glove_w = torch.from_numpy(dataset_val.glove_w).float()
opt.st2towidx = torch.from_numpy(dataset_val.st2towidx).long()
opt.itow = dataset_val.itow
opt.itod = dataset_val.itod
opt.ltow = dataset_val.ltow
opt.itoc = dataset_val.itoc
# choose the attention model
if opt.att_model == 'topdown':
model = AttModel.TopDownModel(opt)
else:
model = AttModel.Att2in2Model(opt)
if opt.start_from is not None:
if opt.load_best_score == 1:
model_path = os.path.join(opt.start_from, 'model-best.pth')
info_path = os.path.join(opt.start_from,
'infos_' + opt.id + '-best.pkl')
else:
model_path = os.path.join(opt.start_from, 'model.pth')
info_path = os.path.join(opt.start_from,
'infos_' + opt.id + '.pkl')
# opt.learning_rate = saved_model_opt.learning_rate
print('Loading the model weights, path is %s...' % (model_path))
model.load_state_dict(torch.load(model_path))
if opt.mGPUs:
model = nn.DataParallel(model)
if opt.cuda:
model.cuda()
####################################################################################
# Evaluate the model
####################################################################################
lang_stats, predictions = eval_NBT(opt,
model,
dataset_val,
processing='eval')
print('print the evaluation:')
for k, v in lang_stats.items():
print('{}:{}'.format(k, v))